Neural Measures of Human Decision Making Track Evidence Accumulation in Learned Space

This study demonstrates that the brain's evidence accumulation mechanism, indexed by the centro-parietal positivity (CPP), is domain-general and flexibly tracks decisions based on learned, arbitrary representational transformations, not just direct sensory or pre-existing evidence.

The Gist

The Big Question: How Does the Brain Decide?

Imagine your brain is a judge in a courtroom. Every time you make a decision, this judge has to weigh the evidence. If the evidence is strong (like a clear video of a crime), the judge makes a decision quickly. If the evidence is weak or blurry (like a witness with bad eyesight), the judge takes longer to think it over.

Scientists have known for a long time that the brain uses a specific "neural signal" to do this weighing. In the brain, this signal is called the CPP (Centro-Parietal Positivity). You can think of the CPP as a filling bucket.

• As evidence comes in, water (information) pours into the bucket.
• The faster the water pours in, the quicker the bucket fills up to the "decision line."
• Once the bucket is full, you make a choice.

The Mystery: Does the Bucket Work for Learned Rules?

Previous studies showed this "filling bucket" works great for:

1. Direct senses: Seeing a blurry dot moving left or right.
2. Memory: Remembering facts (e.g., "Is this state bigger than that one?").
3. Fixed rules: Knowing that "vertical lines are Category A" and "diagonal lines are Category B."

But there was a big gap in our knowledge. What happens when the rule isn't fixed? What if the rule is arbitrary and learned on the spot?

The Analogy:
Imagine you are playing a game where you have to sort red and blue balls.

• Old studies: The rule was always "Red goes left, Blue goes right." Everyone knew this.
• This study: The experimenter says, "Today, the rule is: If the ball is tilted more than 30 degrees to the left, it goes left. If it's tilted less, it goes right." But here's the catch: The 30-degree line is different for every single person. You have to figure out your own specific line through trial and error.

The scientists wanted to know: Does the brain's "filling bucket" still work when the rule is a weird, personal, learned math problem?

The Experiment: The "Tilted Stick" Game

The researchers put 38 people in an EEG cap (a helmet that reads brain waves) and played a game:

1. Training: Participants saw a bunch of sticks tilted at different angles. They had to guess which "category" (Left or Right) the stick belonged to. They got feedback ("Correct!" or "Wrong!") until they figured out their personal "cut-off line."
2. The Test: Once they learned their specific line, the researchers showed them new sticks.
   • Some sticks were far away from the line (e.g., way over to the left). This is strong evidence.
   • Some sticks were very close to the line (e.g., just a tiny bit to the left). This is weak evidence.

The Results: The Bucket is Universal!

The scientists found two amazing things:

1. The Bucket Fills Faster with Stronger Evidence
Just like in the old studies, when the stick was far from the line (strong evidence), the brain's "filling bucket" (the CPP) filled up fast. When the stick was close to the line (weak evidence), the bucket filled up slowly.

• The Takeaway: It didn't matter that the rule was learned and arbitrary. The brain treated the "distance from the line" as real evidence, just like it treats visual clarity.

2. The Brain and the Math Matched
The researchers used a computer model (Drift-Diffusion Model) to calculate how fast each person should be accumulating evidence based on their reaction times. Then, they looked at the actual brain waves.

• The Result: The people whose brains filled the bucket faster (steeper CPP slope) were the exact same people whose computer models showed faster evidence accumulation.
• The Metaphor: It's like checking a car's speedometer (the brain wave) against the GPS calculation (the math model). They matched perfectly, proving the brain is actually doing the math of "evidence accumulation."

What This Means for You

This study proves that the brain's decision-making machinery is incredibly flexible.

Think of the brain's decision center as a universal translator.

• It doesn't care if the information comes from your eyes (senses).
• It doesn't care if it comes from your memory (facts).
• It doesn't care if it comes from a new, weird rule you just learned (like the tilted sticks).

As long as there is a "distance" to measure between what you see and the rule you learned, the brain converts that into a steady stream of evidence, fills up its bucket, and makes a decision.

In short: The brain doesn't just react to the world; it builds its own internal maps, learns new rules on the fly, and uses the same powerful "evidence-filling" engine to navigate them all.

Technical Summary

1. Problem Statement and Research Question

The study addresses a fundamental gap in decision neuroscience regarding the domain-generality of evidence accumulation.

• Background: The "accumulation-to-threshold" framework posits that decisions result from integrating evidence until a threshold is reached. The Centro-Parietal Positivity (CPP), a scalp EEG potential, is the established neural signature of this process. Previous research has shown CPP tracks evidence from sensory inputs, working memory, semantic knowledge, and fixed perceptual axes (e.g., cardinal vs. diagonal orientations).
• The Gap: It remains unknown whether the CPP tracks decisions based on evidence that must be computed through newly learned, arbitrary, and participant-specific representational transformations. In other words, does the brain's decision machinery adapt to evidence defined entirely by a learned rule, rather than innate sensory features or fixed perceptual frameworks?

2. Methodology

Experimental Design

• Participants: 38 human volunteers (after exclusions) aged 18–40.
• Task Structure:
  1. Training Phase: Participants learned to classify continuously-oriented visual stimuli (apertures of iso-oriented bars) into two discrete categories ("Category 1" or "Category 2").
     • Arbitrary Boundary: Each participant was assigned a unique, random category boundary (an orientation angle between 0° and 179°).
     • Learning: Participants identified this boundary through trial-and-error with feedback.
  2. Main Experiment: Participants performed the classification task while undergoing EEG recording.
     • Stimuli: Oriented bars presented for 3 seconds.
     • Conditions: Stimuli were presented at varying angular distances from the learned boundary (Category Coherence): ±2°, ±5°, ±15°, and ±45°.
     • Metric: "Category Coherence" was operationalized as the angular distance between the stimulus and the learned boundary.

Data Acquisition and Preprocessing

• EEG: Recorded from 63 scalp electrodes (BrainProducts actiCHamp Plus) at 1 kHz, resampled to 250 Hz.
• Preprocessing: Included high-pass filtering (0.5 Hz), artifact subspace reconstruction, interpolation of bad channels, re-referencing to the mean, and application of a surface Laplacian to minimize spatial cross-contamination.
• CPP Definition: Defined as the average voltage over centro-parietal electrodes (CP1, CPz, CP2, P1, Pz, P2).
• Metric: The CPP buildup rate was calculated as the slope of the response-locked waveform from -1.0s to 0.0s relative to response onset.

Computational Modeling

• Drift-Diffusion Model (DDM): A DDM was fitted to each participant's accuracy and reaction time (RT) data using the pyDDM framework.
• Parameterization:
  • Drift Rate ($v$): Allowed to vary as a function of category coherence.
  • Decision Bound ($a$) & Non-decision Time ($T_{er}$): Held constant across coherence levels.
• Model Selection: Bayesian Information Criterion (BIC) was used to compare the constrained model (M1) against models where bounds or non-decision times varied.

Additional Analyses

• Correlation: Pearson correlation was computed between individual differences in coherence-drift rate slopes (behavioral) and coherence-CPP slopes (neural).
• Motor Control: Lateralized beta-band power (15–25 Hz) over motor cortex (C3/4, FC3/4) was analyzed to rule out motor preparation as the source of CPP modulations.

3. Key Results

Behavioral Performance

• Participants learned the arbitrary categorization rule rapidly (asymptotic performance by block 2–3).
• Accuracy and RT: Both improved monotonically with increasing category coherence (stimuli farther from the boundary were classified faster and more accurately).
• DDM Fitting: The model provided an excellent fit ($R^2 > 0.99$ for accuracy and RT). Drift rates increased significantly with category coherence, confirming that difficulty was driven by evidence strength. Model M1 (varying drift only) was favored over more complex models for 34/38 participants.

Neural Findings (CPP)

• CPP Buildup: Response-locked CPP waveforms showed a ramping buildup prior to response.
• Coherence Scaling: The CPP buildup rate (slope) increased monotonically with category coherence. High-coherence trials (far from boundary) exhibited steeper slopes than low-coherence trials.
• Neural-Behavioral Coupling: There was a significant positive correlation ($r = 0.50, p < 10^{-4}$) across participants between the sensitivity of the drift rate to coherence and the sensitivity of the CPP slope to coherence. Individuals with steeper behavioral drift-rate functions also showed steeper neural accumulation slopes.

Motor Specificity

• Lateralized Beta Power: While beta power showed the expected contralateral suppression (motor preparation) prior to response, the slope of this suppression did not vary with category coherence ($BF_{01} = 36.76$, strong evidence for the null).
• Conclusion: The CPP modulation is distinct from effector-specific motor preparation and reflects central evidence accumulation.

4. Key Contributions

1. Extension of Domain-Generality: The study provides the first direct evidence that the CPP tracks evidence accumulation in learned, arbitrary category spaces. This confirms that the neural decision machinery is not limited to sensory evidence or fixed perceptual axes but adapts to internally constructed, task-specific representational frameworks.
2. Validation of CPP as a Decision Variable: The strong correlation between individual differences in computational drift rates and neural CPP slopes validates the CPP as a direct neural index of the decision variable, even when that variable is computed via complex learned transformations.
3. Dissociation from Motor Output: By demonstrating that motor beta slopes are insensitive to evidence strength while CPP slopes are sensitive, the study reinforces the interpretation of CPP as a central, effector-independent decision signal.
4. Cross-Species Convergence: The findings align with macaque single-unit data showing that Lateral Intraparietal (LIP) neurons encode learned categorical boundaries, suggesting a conserved parietal mechanism for domain-general evidence accumulation across species.

5. Significance

This research fundamentally advances the understanding of human decision-making by demonstrating that the brain's accumulation mechanism is representationaly flexible. The CPP does not merely reflect the integration of raw sensory input; it integrates task-relevant evidence regardless of its origin—whether derived from direct perception, memory retrieval, or learned computational transformations.

This implies that the neural circuits governing decision-making (likely involving parietal cortex) operate on abstract decision variables. The brain can dynamically map sensory features onto internal, learned coordinate systems and accumulate evidence within those systems using the same universal mechanism used for simple perceptual discrimination. This has profound implications for theories of cognitive flexibility, rule-based learning, and the neural architecture of complex categorization.